J. Chem. Inf. Comput. Sci. 1999, 39, 133-138
133
Wavelet Neural Network and Its Application to the Inclusion of β-Cyclodextrin with Benzene Derivatives Lei Liu and Qing-Xiang Guo* Department of Chemistry, University of Science and Technology of China, Hefei, 230026, People’s Republic of China Received June 9, 1998
A wavelet neural network (WNN) was constituted and applied to the inclusion complexation of β-cyclodextrin with mono- and 1,4-disubstituted benzenes. The association constant (Ka) values have been calculated by the WNN from substituent molar refraction (Rm), hydrophobic constant (π), and Hammett constant (σ) of the guest compounds as input parameters. The excellent prediction results with a correlation coefficient of 0.992 and standard deviation of 0.089 suggested that β-CD inclusion complexation is mainly driven by van der Waals force, hydrophobic interaction, and electronic effects. INTRODUCTION
R-, β-, and γ-cyclodextrin (CD) are cyclic oligomers of six, seven, and eight D-glucose units. They can form inclusion complexes with a variety of guest molecules through a process called molecular recognition. This property enables CD to be applied to many important areas, such as analytical chemistry, pharmaceutical science, catalysis, and separation technology.1 CD inclusion complexes are also the most valuable models for understanding noncovalent bonding interactions in aqueous solution and mimicking the enzyme-substrate interaction.2 Great efforts have been devoted to the driving forces of CD inclusion complexation. Up to date, several driving forces have been postulated in the host-guest interaction: (1) van der Waals forces; (2) hydrophobic interactions; (3) hydrogen bonding; (4) release of distortional energy of CD by binding guest; and (5) extrusion of “high energy water” from the cavity of CD upon inclusion complex formation. However, there still remains no clear agreement on the mechanism for the CD inclusion complexation.3a Recently, quantitative studies on the CD inclusion complexation have attracted much attention with the purposes of both understanding the roles of different driving forces and theoretically predicting the association constants (Ka) of CD complexes. Methods including quantum and molecular mechanics computation4 and linear regression5 as well as artificial neural networks (ANN)6 have been used. Many studies have demonstrated that there exists quantitative relation between CD association constants and the substituent properties such as volume, hydrophobicity, and electronic properties of the guest molecules. Although some reliable prediction models have been established for R-CD complexes,3b,5 little work has been performed on β-CD complexation systems. A linear model for the β-CD binding with 1,4-disubstituted benzenes has been reported, but the correlation coefficient of only 0.723 could be reached for a rather small sample of 16 complexes.5a In our previous work, a multiple linear regression model and a stepwise regression analysis model for the binding of β-CD with 24 monosubstituted benzenes were established with correlation coefficients of 0.917 and 0.94,8 respectively. Our approach was
also well applied to the β-CD‚1-substituted naphthalene system.9 However, to obtain a better model for prediction, some nonlinear analysis method is needed because of the advent of nonlinear mechanisms in the process caused by the interactions of the different driving forces. It is well known that the artificial neural network is a high performance nonlinear analysis tool, which is capable of covering more relations between the input/output data than classical regression analysis without prior knowledge of the relationship between variables involved in the system.10 Although great interests have recently been devoted to the application of ANN to different domains of chemistry, potential applications of ANN in supramolecular chemistry still require further exploration. Recently, we have reported the prediction of the driving forces for R-CD complexation with benzene derivatives by the BP networks.6 In this study, a novel wavelet neural network (WNN) was employed in the inclusion complexation of β-CD with mono- and 1,4disubstituted benzenes, and an excellent driving force prediction was achieved.
THEORY
1. Wavelet Neural Network. Wavelet theory, a novel field which interests not only mathematicians but also scientists studying acoustics, fluid mechanics, and chemistry, involves representing general signals in terms of simpler, fixed building blocks of constant shape but at different scales and positions. Unlike Fourier transformation, the wavelet transformation has dual localization both in frequency and in time, which makes it an excellent nonlinear analysis tool.11
10.1021/ci980097x CCC: $18.00 © 1999 American Chemical Society Published on Web 10/07/1998
134 J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999
LIU
( )
tn′2 tn′ ∂h ) -cos(1.75tn′) exp ∂xn′ 2 at
AND
( )
1.75 sin(1.75tn′) exp -
( )
tn′2 tn′2 ∂h ) cos(1.75tn′) exp + ∂at 2 at
Recently, wavelet neural networks (WNN) have been proposed based on the theories of feed-forward neural networks and wavelet decompositions.12 Although several theoretical studies have proven the exciting superiority of WNN over conventional BP neural network,12 very little has been reported on the applications of WNN, especially in chemistry.13 The structure of WNN employed in this study is shown in Figure 1. The computer program was written in Borland C++ 3.1 and run on an 80586 personal computer; the back-propagation algorithm14 was used in the network training process, which is summarized as follows: (1) Initializing dilation parameter at, translation parameter bt, and node connection weights utiwt to some random values. (2) Inputting data Xn(i) and corresponding output values VTn , where superscript T represents the target output state. (3) Propagating the initial signal forward through the network using
( ) S
T
Vn ) ∑wth t)1
∑utixn(i) - bt
( )
tn′2 tn′ ∂h ) cos(1.75tn′) exp + ∂bt 2 at
tn′2 1 2 at
( )
1.75 sin(1.75tn′) exp Figure 1. The architecture of the wavelet neural network.
GUO
tn′2 tn′ 2 at
( )
1.75 sin(1.75tn′) exp -
tn′2 1 2 at
and the objective function E is taken as
E)
1
N
(VTn - Vn)2 ∑ 2 n)1
(5) Using
∂E ∂E old ∆wnew ) -η old + R∆wold ∆unew t t ti ) -η old + R∆uti ∂wt ∂uti ∂E ∂E ∆anew ) -η old + R∆aold ∆bnew ) -η old + R∆bold t t t t ∂at ∂bt updating at, bt, wt, uti to
i)1
old old old unew anew ) aold ti ) uti + ∆uti t t + ∆at
at
where h is taken as a Morlet wavelet
old old bnew ) bold wnew ) wold t t + ∆bt t t + ∆wt
( )
t2 2 (4) Calculating the variation of every parameter in the network, defined as h(t) ) cos(1.75t) exp -
( )
where η is the learning rate and R the momentum. (6) Returning to step 2 and repeating for another round of training. This process is continued until the network output satisfies the error criteria. 2. Parameters.15 Molecular refraction is defined as
S
δwt )
∂E ∂wt
N
)-
n)1
δat ) δbt ) δuti )
∑ (VTn - Vn)h
∂E ∂at ∂E ∂bt
∂E ∂uti
N
)-
∑utixn(i) - bt
Rm ) [(n2 - 1)/(n2 + 2)]‚MW/d
at
where n, MW, and d represent the refractivity, molecular weight, and density of the compound, respectively. In this work the substituent molar refraction was used, in which MW is replaced by the weight of the substituent. Substituent hydrophobic constant πx is defined as
i)1
∂
∑ (VTn - Vn)wt∂a
n)1 N
)-
)-
∂h
∑ (VTn - Vn)wt∂b
n)1 N
t
t
∂h
∑ (VTn - Vn)wt∂x ′xn(i)
n)1
n
S utixn(i) and tn′ ) (xn′ - bt)/at for convenwhere xn′ ) ∑i)1 ience,
πx ) log PX - log PH where PX and PH are the partition coefficients of substituted compound (RX) and the parent compound (RH), such as PhX and PhH, in 1-octanol/water. Hammett σ constant, the electronic effect of substituent (X) relative to that of hydrogen, may be obtained by a comparison of ∆G for dissociation constants of substituted benzoic acid (KX) with that of the parent compound, benzoic
WAVELET NEURAL NETWORK
J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999 135
Table 1. ln Ka Values Calculated by the WNN and the Experimental Data for the Inclusion Complexation of β-CD with Mono- and 1,4-Disubstituted Benzenes no.
X
Y
ln Ka (obs)
ref
ln Ka (WNN)
RmX
πX
σX
RmY
πY
σY
1 2a 3 4 5 6a 7 8 9a 10 11 12a 13 14 15a 16 17 18 19 20 21a 22 23 24 25a 26 27 28 29 30 31 32 33 34 35 36 37a 38 39 40
H CH3 Et CCH OH OCH3 OEt CH2OH CH2Cl CHO COMe CO2Me CO2Et CN NHCH3 NHEt N(CH3) 2 NHCOMe NO2 F Br I NH2 CH3 Br I CH2OH Et NO2 COOH NO2 I CH3CO Br CH3O OH Cl CH3 COOH Cl
H H H H H H H H H H H H H H H H H H H H H H H CH3 Br I OH OH OH OH NH2 OH OH OH OH OH OH OH NO2 NO2
5.13 5.37 5.96 5.44 4.55 5.34 5.73 4.96 5.63 5.01 5.23 5.76 6.29 5.14 4.87 5.38 5.44 5.06 5.63 4.51 5.76 6.74 3.91 5.48 6.85 7.31 4.98 6.20 5.50 5.06 5.72 6.86 5.02 6.10 5.09 4.73 6.02 5.52 5.39 4.95
17 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 18 19 20 20 21 21 22 22 22 23 24 24 24 25 25 25 25 26
5.13 5.40 5.96 5.42 4.46 5.29 5.67 4.86 5.58 5.16 5.19 5.58 6.38 5.09 4.67 5.42 5.36 5.04 5.68 4.55 5.89 6.69 3.96 5.48 6.87 7.28 4.97 6.20 5.34 5.22 5.74 6.90 5.04 6.05 5.09 4.77 6.32 5.50 5.38 4.96
0.34 5.07 9.77 8.52 5.05 9.44 13.86 9.31 14.11 8.78 13.00 16.42 20.59 7.79 9.96 14.69 14.48 15.04 12.29 5.17 17.25 24.38 5.25 5.07 17.25 24.38 9.31 9.77 12.29 13.07 12.29 24.38 13.00 17.25 9.44 5.05 9.83 5.07 13.07 9.83
0.00 0.56 1.02 0.40 -0.67 -0.02 0.44 -1.03 0.17 -0.65 -0.55 -0.01 0.45 -0.57 -0.36 0.13 0.18 -0.97 -0.28 0.14 0.86 1.12 -1.23 0.56 0.86 1.12 -1.03 1.02 -0.28 -0.28 -0.28 1.12 -0.55 0.86 -0.02 -0.67 0.71 0.56 -0.28 0.71
0.00 -0.17 -0.15 0.23 -0.37 -0.32 -0.24 0.08 0.18 0.22 0.50 0.39 0.45 0.66 -0.84 -0.61 -0.83 0.00 0.78 0.06 0.27 0.30 -0.66 -0.17 0.27 0.30 0.08 -0.15 0.78 0.45 0.78 0.30 0.50 0.27 -0.32 -0.37 0.30 -0.17 0.45 0.30
0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 0.34 5.07 17.25 24.38 5.05 5.05 5.05 5.05 5.25 5.05 5.05 5.05 5.05 5.05 5.05 5.05 12.29 12.29
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.56 0.86 1.12 -0.67 -0.67 -0.67 -0.67 -1.23 -0.67 -0.67 -0.67 -0.67 -0.67 -0.67 -0.67 -0.28 -0.28
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.17 0.27 0.30 -0.37 -0.37 -0.37 -0.37 -0.66 -0.37 -0.37 -0.37 -0.37 -0.37 -0.37 -0.37 0.78 0.78
a
Used in the prediction set.
acid (KH).
σX ) ∆GX - ∆GH ) log KX - log KH 3. Orientation of the Guest Compound in the Complex. The determination of the orientation of the guest molecule in CD cavity is a rather complicated and controversial problem not only theoretically5 but also experimentally.16 In our present study, the orientation of the benzene derivatives was postulated as following: 1. For monosubstituted benzene, since the substituent groups are generally far larger than hydrogen, which in consequence fit more snugly in β-CD cavity Via van der Waals forces, the substituent groups are proposed to stay in CD cavity.7 2. For para-substituted phenol and para-substituted aniline, the OH and NH2 are readily proposed to stay outside since OH and NH2 are highly hydrophilic. 3. For p-nitrobenzoic acid, the carboxyl group was proposed to stay in CD cavity, since the volume of the COOH is larger than that of NO2. 4. For p-chloronitrobenzene, the nitro group was proposed to stay outside the β-CD cavity, because Cl is more hydrophobic than NO2.
In this paper, the substituent X is located in the β-CD cavity and the substituent Y outside the cavity. Despite the empirical nature of the above postulations, the excellent prediction of the WNN gives further confidence on them. CALCULATION
The association constants (ln Ka) and the corresponding substituent constants Rm, π, and σ of 32 benzene derivatives were used to train the network. After the network training, to examine the network eight other association constants were predicted. Different 6-n-1 WNNs (n ) 1-10) were also respectively trained. It was found that the WNNs with both very few and too many hidden nodes have difficulties in quickly generalizing the model. In this study, the topological structure capable of the best performance and hence selected was found to be 6-6-1. Several independent networks of the 6-6-1 structure were also trained respectively in order to avoid the possible chance effects.14b However, since no significant difference in the network predictions between them was observed in our study, we here only list the predictions of one network in Table 1, with its optimized weights summarized in Table 2.
136 J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999
LIU
AND
GUO
Table 2. Optimized Weights of the Wavelet Neural Network hidden node
1
2
3
4
5
6
U1 U2 U3 U4 U5 U6 a b w
-2.888 110 -8.561 712 -1.997 538 -0.036 586 4.643 066 0.401 329 -54.446 556 1.398 565 -2.276 946
5.162 836 2.137 373 3.611 767 6.326 932 -1.190 588 1.016 571 -32.699 768 -0.273 823 -1.837 876
-5.036 614 5.826 611 -4.680 911 -1.889 969 -2.418 538 -1.272 372 -23.379 061 -6.062 471 4.742 377
-0.968 686 -0.866 055 -1.887 442 1.083 224 -0.319 131 -0.149 267 145.266 556 0.716 122 6.615 698
4.218 651 -8.232 189 -4.597 124 1.734 114 4.381 264 2.841 160 41.205 219 -8.888 271 7.015 137
5.185 442 -9.754 425 -5.876 666 2.649 754 6.686 687 2.372 448 35.503 597 0.956 005 -8.035 176
Figure 2. Plot of ln Ka (obs) versus ln Ka(WNN) for the inclusion complexation of β-CD with benzene derivatives.
From Table 1 it can be seen that the ln Ka (WNN) values predicted by WNN from the substituent Rm, π and σ constants are very close to those determined experimentally. Plotting ln Ka observed against those predicted by WNN gives a straight line (Figure 2), which fits the following equation:
ln Ka(obs) ) 0.976 ln Ka(WNN) + 0.134
(1)
n ) 40, r ) 0.992, sd ) 0.089 Although the calculation sample in our study is somewhat small, due to the limited experimental data in the literature, the performance of the wavelet neural network demonstrates that WNN is a promising novel nonlinear analysis tool. DISCUSSION
1. WNN versus BP Neural Networks. It is worthy to note that WNNs replace the Sigmoid transfer function in conventional BP neural networks with scalar wavelets. However, this replacement results in nontrivial improvement due to the desirable properties of wavelet transformation. Theoretical studies as well as simulation results have confirmed the superiority of WNN.12 Herein, in order to compare the performance of WNN with BP neural network based on the real physical model in our study, a BP network
of 6-6-1 structure was also implemented.27 This BP network yields predictions for the same sample with a correlation coefficient of 0.981. It is obvious that WNN can offer a better estimation due to the desirable properties of wavelet transformation. This result is in agreement with previous studies.12,13 It is therefore hopeful that the novel wavelet neural network could be a good alternative of the BP network, which has already been heavily used in chemistry. 2. Driving Forces for CD Inclusion Complexation. The above WNN model demonstrates that β-CD inclusion complexation is affected by substituent molar refraction (Rm) of the guest molecules. Since Rm well reflects the volume and polarizability of the substrate, it readily concluded that CD host-guest complexation is affected by van der Waals forces. The van der Waals forces primarily consist of induction and dispersion forces, which depend on molecular volume and polarizability. Many studies have confirmed that van der Waals forces play overwhelming roles in CD inclusion complexation.28 Another factor influencing the CD inclusion complexation is hydrophobic interaction,29 which was included in above WNN model by hydrophobic constant (π). The involvement of hydrophobic interaction in CD inclusion complexation has been disputable for years. However, many authors tend to conclude that hydrophobic interaction plays at least a significant role, if not a dominant role, in CD inclusion
WAVELET NEURAL NETWORK
complexation. Although the hydrophobic interaction partly results from the van der Waals force, it is mainly due to the effects of entropy produced in the water molecules. The substituent with larger π values is more hydrophobic and thus strongly driven into the hydrophobic cavity of R-CD from the water cluster. This process is exothermic by entropic gain.7 Since CD is highly polar,30a electrostatic force, sometimes also referred as dipole-dipole interaction, can to some extent contribute to the CD inclusion complexation.30 However, ambiguity exists because this type of interaction is often combined with dipole-induced dipole and dispersion forces. Different authors ranked electrostatic force in CD inclusion complexation extremely differently. The substituent Hammett σ constant in above WNN model well reflects this factor. Although still other factors may contribute to the CD inclusion mechanism,3a,31 the successful prediction of above WNN model suggests that van der Waals forces, hydrophobic interaction, and electronic effect comprise the major driving forces. 3. Multilinear Regression versus Neural Network. To be compared with the performance of the WNN, the multilinear regression (MLR) analysis is implemented for the above 40 inclusion complexes with Rm, π, and σ constants as input descriptors. The optimized MLR equation is obtained as follows:
ln Ka ) 4.779 + 0.050RmX + 0.539πX + 0.273σX + 0.019RmY + 0.349πY - 0.641σY (2) r2 ) 0.893, sd ) 0.244, n ) 40 It is notably better than that reported in the literature5a in which the effect of the hydrophobic interaction was neglected. Omitting πX and πY from eq 2 could only yield ln Ka ) 4.382 + 0.081RmX + 0.081σX + 0.037RmY - 0.332σY (3) r2 ) 0.638, sd ) 0.437, n ) 40 From eq 2, it can be seen that the larger the Rm and π constants of the substituents, the greater the stability of the inclusion complexes. This indicates that stronger van der Waals forces and hydrophobic interactions favor the molecular recognition process.7 It is also interesting to notice that differently located substituents play different roles in electronic interactions according to eq 2. The group (X) inside β-CD cavity positively contributes to the binding, while the group (Y) outside CD cavity negatively affects the binding if it possesses more positive σ constant. This agrees with other studies which proposed that the antiparallel arrangement of the polarity of the guest with CD favors the binding.5a,30 Although above MLR model offers more insight into the inclusion mechanism, it is obvious that WNN covers more relations than conventional regression and therefore gives much better predictions. Hence, the WNN is more recommendable for the purpose of prediction association constants of CD complexation. CONCLUSIONS
A novel wavelet neural network was successfully employed to predict the inclusion of β-CD with mono- and 1,4-
J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999 137
disubstituted benzenes from substituent molar refraction Rm, hydrophobic constant π, and Hammett constant σ. The study well supported the postulation that van der Waals forces, hydrophobic interactions, and electronic effects mainly contribute to the β-CD inclusion complexation. The study also demonstrates the desirable properties of the WNN surely deserve further exploration. ACKNOWLEDGMENT
This research was supported by the National Natural Science Foundation of China. REFERENCES AND NOTES (1) (a) Bender, M. L.; Komiyama, M. Cyclodextrin Chemistry; SpringerVerlag: Berlin, 1978. (b) Szejtli, J. Cyclodextrin Technology; Kluwer Academic Publishers: Dordrecht, 1988. (2) Breslow, R. Biomimetic chemistry and artificial enzymes catalysis by design. Acc. Chem. Res. 1995, 28, 146-153. (3) (a) Connors, K. A. The stability of cyclodextrin complexes in solution. Chem. ReV. 1997, 97, 1325-1358. (b) Connors, K. A. Prediction of bind constants of R-cyclodextrin complexes. J. Pharm. Sci. 1996, 85, 796-802. (4) (a) Pe´rez, F.; Jaime, C.; Sa´nchez-Ruiz, X. MM2 calculations on cyclodextrins: Multimodel inclusion complexes. J. Org. Chem. 1995, 60, 3840-3845. (b) Huang, M. J.; Watts, J. D.; Bodor, N. Theoretical studies of inclusion complexes of R- and β-cyclodextrin with benzoic acid and phenol. Inter. J. Quan. Chem. 1997, 65, 1135-1152. (5) (a) Davies, D. M.; Savage, J. R. Correlation analysis of the host-guest interaction of R-cyclodextrin and substituentd benzenes. J. Chem. Res. (S) 1993, 94-95; J. Chem. Res. (M) 1993, 660-671. (b) Park, J. H.; Nah, T. H. Binding forces contributing to the complexation of organic molecules with β-cyclodextrin in aqueous solution. J. Chem. Soc., Perkin Trans. 2 1994, 1359-1362. (6) (a) Guo, Q.-X.; Luo, S.-H.; Wang, H.; Zhang, M.-S.; Liu, Y.-C. A novel approach to the prediction of binding constants for the inclusion complexation of R-cyclodextrins with benzene derivatives by artificial neural networks. J. Chem. Res. (S) 1996, 38-39. (b) Guo, Q.-X.; Chu, S.-D.; Wang, Y.-M.; Pan, Z.-X.; Zhang, M.-S.; Liu, Y.-C. A study on the inclusion of R-cyclodextrins with mono- and 1,4-disubstituted benzenes by neural networks. Chin. Chem. Lett. 1997, 8, 145-148. (7) Guo, Q.-X.; Luo, S.-H.; Liu, Y.-C. Substituent effects on the driving force for inclusion complexation of R- and β-cyclodextrins with monosubstituted benzene derivatives. J. Incl. Phenom. 1998, 30, 173-182. (8) (a) Guo, Q.-X.; Luo, S.-H.; Zheng, X.-Q.; Liu, Y.-C. A regression study on the inclusion of cyclodextrin with benzene derivatives. Chin. Chem. Lett. 1996, 7, 776-770. (b) Zhang, H.-M.; Luo, S.-H.; Chen, C.; Liu, L.; Guo, Q.-X.; Liu, Y.-C. Studies on the driving force for inclusion complexation of R- and β-cyclodextrin with substituted benzenes. Chem. Res. Chin. UniV., in press. (9) Guo, Q.-X.; Zheng, X.-Q.; Ruan, X.-Q.; Luo, S.-H.; Liu, Y.-C. Substituent effects and enthalpy-entropy compensation on the inclusion of β-cyclodextrins with 1-substituted naphthalenes. J. Incl. Phenom. 1996, 26, 175-183. (10) (a) Gasteiger, J.; Zupan, J. Neural networks in chemistry. Angew. Chem., Int. Ed. Engl. 1993, 32, 503-527. (b) Burns, S. A.; Whitesides, G. M. Feed-forward neural networks in chemistry: mathematical systems for classification and pattern recognition. Chem. ReV. 1993, 93, 2583-2601. (c) Sumpter, B.; Getino, C.; Noid, D. Theory and applications of neural computing in chemical science. Ann. ReV. Phys. Chem. 1994, 45, 439-481. (11) Chui, K. An introduction to waVelets; Academic Press: San Diego, 1992. (12) (a) Zhang, Q.-H.; Benveniste, A. Wavelet networks. IEEE Trans. Neural Network 1992, 23, 889-898. (b) Pati, Y. C.; Krishnaprasad, P. S. Analysis and synthesis of feed-forward neural networks using discrete affine wavelet transformations. IEEE Trans. Neural Network 1993, 4, 73-87. (c) Zhang, J.; Walter, G. G.; Miao, Y.; Lee, W. N. Wavelet neural networks for function learning. IEEE Trans. Signal Processing 1995, 43, 1485-1496. (d) Zhang, Q.-H. Using wavelet network in non-parametric estimation. IEEE Trans. Neural Network 1997, 8, 227-236. (13) (a) Qian, W.; Clarke, L. P. restoration algorithm for P-32 and Y-90 bremsstrahlung emission nuclear imaging: a wavelet-neural network approach. Med. Phys. 1996, 23, 1309-1323. (b) Qing, W.; Shen, X.Q.; Yang, Q.-X.; Yan, W.-L. Using wavelet neural networks for the optimal design of electromagnetic devices. IEEE Trans. Magnetics 1997, 33, 1928-1930.
138 J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999 (14) (a) Rumelhart, D. E.; Hinton, G. E.; Williams, R. J. Learning representations by back-propagating errors. Nature 1986, 323, 533536. (b) Livingstone, D. J.; Manallack, D. T. Statistics using neural networks: chance effects. J. Med. Chem. 1993, 36, 1295-1297. (15) (a) Hansch C.; Leo, A.; Taft, R. W. Chem. ReV. 1991, 91, 165. (b) Hansch, C.; Leo, A. Substituent constants for correlation analysis in chemistry and biology; Wiley: New York, 1979. (16) Alderfer, J. L.; Eliseev, A. V. Complex of 4-flurophenol with R-cyclodextrin: binding mode in solution is opposite to that in solid state. J. Org. Chem. 1997, 62, 8225-8226. (17) Tucker, E. E.; Christian, S. D. Vapor pressure studies of benzenecyclodextrin inclusion comlexes in aqueous solution. J. Am. Chem. Soc. 1984, 106, 1942-1945. (18) Hoshino, M.; Imamura, M.; Ikehara, K.; Hama Y. Fluorescence enhancement of benzene derivatives by forming inclusion complexes with β-cyclodextrins in aqueous solutions. J. Phys. Chem. 1981, 85, 1820-1823. (19) Sanemasa, I.; Akamine, Y. Association of symmetrical p-dihalobenzenes with cyclodextrins in aqueous medium. Bull. Chem. Soc. Jpn. 1987, 60, 2049-2066. (20) Takuma, T.; Deguchi, T.; Sanemasa, I. Association of symmetrical p-dihalobenzenes with cyclodextrins in aqueous medium. Bull. Chem. Soc. Jpn. 1991, 64, 480-484. (21) Kano, K.; Tamiya, Y.; Hashimoto, S. Binding forces in complexation of p-Alkylphenols with β-cyclodextrin and methylated β-cyclodextrins. J. Incl. Phenom. 1992, 13, 287-293. (22) Shimizu, H.; Kaito, A.; Hatano, M. Induced circular dichroism of β-cyclodextrin complexes with substituted benzenes. Bull. Chem. Soc. Jpn. 1979, 52, 2678-2684. (23) Rudiger, V.; Eliseev, A.; Simova, S.; Schneider, H.-J.; Blandamer, M. J.; Cullis, P. M.; Meyer, A. J. Conformational, calorimetric and NMR spectroscopic studies on inclusion complexes of cyclodextrins with substituted phenyl and adamantane derivatives. J. Chem. Soc., Perkin Trans. 2 1996, 2119.
LIU
AND
GUO
(24) Liu, Y.; You, C.-C.; Han, B.-H.; Li, Y.-M.; Li, B.; Chen, R.-T. Molecular recognition study on supermolecular system (XII). Chem. J. Chin. UniV. 1997, 18, 1316-1320. (25) Bertrand, G. L.; Faulkner, J. R.; Han, S. M.; Armstrong, D. W. Substituent effects on the binding of phenols to cyclodextrins in aqueous solution. J. Phys. Chem. 1989, 93, 6863. (26) Taraszewska, J.; Piasecki, A. K. Inclusion complexes of isomeric chloronitrobenzenes with R- and β-cyclodextrins studied by polarography analysis of the possibilities of the method. J. Electronal. Chem. 1987, 226, 137. (27) Gasteiger, J.; Zupan, J. Neural networks for chemists: an introduction; VCH Publisher: Weinheim, 1993. (28) (a) Tabushi, I.; Kiyosuke, Y.; Sugimoto, T.; Yamamura, K. Approach to the aspects of driving force of inclusion by R-cyclodextrins. J. Am. Chem. Soc. 1978, 100, 916-919. (b) Tabushi, I. Cyclodextrin catalysis as a model for enzyme action. Acc. Chem. Res. 1982, 66-72. (29) (a) Eftink, M. R.; Andy, M. L.; Bystrom, K.; Perlmutter, H. D.; Kristol, D. S. Cyclodextrin inclusion complexes: studies of the variation in the size of alicyclic guests. J. Am. Chem. Soc. 1989, 111, 6765-6772. (b) Ross, P. D.; Rekharsky, M. V. Thermodynamics of hydrogen bond and hydrophobic interactions in cyclodextrin complexes. Biophys. J. 1996, 71, 2144-2154. (30) (a) Sakurai, M.; Kitagawa, M.; Hoshi, H.; Inoue, Y.; Chujo, R. A molecular orbital study of cyclodextrin inclusion complexes III. dipole moments of cyclodextrins in various types of inclusion complexes. Carbohydr. Res. 1990, 198, 181-191. (b) Connors, K. A.; Pendergast, D. D. Microscopic binding constants in cyclodextrin systems: complexation of R-cyclodextrin of sym-1,4-disubstituted benzenes. J. Am. Chem. Soc. 1984, 106, 7607-7614. (31) Schneider, H.-J. Mechanism of molecular recognition: investigations of organic host-guest complexes. Angew. Chem., Int. Ed. Engl. 1991, 30, 1417-1436.
CI980097X